SkMTEB is the first comprehensive text embedding benchmark for Slovak, and vocabulary-trimmed E5 adaptations achieve competitive performance with much smaller models.
hub Mixed citations
EmbeddingGemma: Powerful and Lightweight Text Representations
Mixed citation behavior. Most common role is background (38%).
56
Pith papers citing it
Background
38% of classified citations
abstract
We introduce EmbeddingGemma, a new lightweight, open text embedding model based on the Gemma 3 language model family. Our innovative training recipe strategically captures knowledge from larger models via encoder-decoder initialization and geometric embedding distillation. We improve model robustness and expressiveness with a spread-out regularizer, and ensure generalizability by merging checkpoints from varied, optimized mixtures. Evaluated on the Massive Text Embedding Benchmark (MTEB) across multilingual, English, and code domains, EmbeddingGemma (300M) achieves state-of-the-art results. Notably, it outperforms prior top models, both proprietary and open, with fewer than 500M parameters, and provides performance comparable to models double its size, offering an exceptional performance-to-cost ratio. Remarkably, this lead persists when quantizing model weights or truncating embedding outputs. This makes EmbeddingGemma particularly well-suited for low-latency and high-throughput use cases such as on-device applications. We provide ablation studies exploring our key design choices. We release EmbeddingGemma to the community to promote further research.
hub tools
citation-role summary
background 4
baseline 2
method 2
citation-polarity summary
representative citing papers
ContinuousBench shows non-private synthetic text transfers corpus-specific capabilities while state-of-the-art DP methods fail to do so even at ε=100.
Dimension d = O(m^{-2} log n) nearly achieves the optimal margin m^rd(+∞, A) for retrieval embeddings, with matching lower bounds showing d = O(k log(n/k)) suffices and is necessary for m = Θ(k^{-1/2}) on k-sparse query matrices.
BoLT is a benchmark of surrogate models fitted to real LLM experiment data that enables evaluation of Bayesian and black-box optimization methods on multi-fidelity, multi-objective, high-dimensional LLM tasks.
ALEE generates AMR-based English minimal pairs with fine-grained semantic shifts, translates them, and evaluates embedding models on 275+ languages to expose cross-lingual gaps linked to training data and tokenization.
TIGER turns the low-rank attention gradient subspace into a differentiable objective for continuous embedding optimization, improving reconstruction quality and robustness over prior discrete token tests especially under noise or DP.
LEDGER provides a corpus of 4,999 annual reports with 31 labeled KPIs and three benchmarks for page-level retrieval, needle-in-haystack lookup, and full KPI extraction from long documents.
EBA clusters sampled LLM generations in representation space to estimate agreement, outperforming random selection with stable scaling and showing that central positions correlate with higher generation quality.
ICL with LLMs reduces absolute imputation error for survey data versus MICE PMM across MCAR/MAR/MNAR mechanisms and yields narrower intervals with near-nominal coverage.
SEA-Embedding is a fully open text embedding pipeline for Southeast Asian languages that achieves state-of-the-art performance on the SEA-BED benchmark by analyzing data composition, training objectives, and base encoder choices.
TeleEmbedBench is the first multi-corpus benchmark showing LLM-based embedding models significantly outperform traditional sentence-transformers on telecommunications specifications and code for retrieval accuracy and noise robustness.
Reinforcement learning with a multi-part reward teaches LLMs to output independent, meaning-preserving sentence edits that raise argument appropriateness close to full rewriting.
KITScenes LongTail supplies multimodal driving data and multilingual expert reasoning traces to benchmark models on rare scenarios beyond basic safety metrics.
Spectral Tempering derives an adaptive scaling factor γ(k) from the embedding eigenspectrum via local SNR analysis and knee-point normalization to achieve near-optimal compression without training or validation.
UWE is a task-agnostic bi-encoder that uses many-to-many InfoNCE and token-level soft late interaction to achieve zero-shot ranking across unseen work-related target spaces while using far fewer parameters than Qwen3-8B and improving MAP by 4.4 points.
ProjAgent introduces procedural similarity—retrieving code with matching computational logic—via LLM hidden-state projections, improving repository-level code generation to 41.14% Pass@1 on REPOCOD.
Graph-PRefLexOR fine-tunes graph-native models with GRPO to organize reasoning into phases, yielding 40-65% gains in traceable hypothesis generation and 2-3x semantic diversity on 100 materials science questions.
Prompt injection defenses create a security-fidelity tradeoff with no model or defense achieving both high security and high fidelity on the SecFid benchmark across 1,168 examples.
Bandit algorithms learn optimal jailbreaks from noisy exploration and, paired with complexity-enhanced queries in FrankensteinBench, achieve up to 97% attack success on 15 open-weight LLMs.
SHIFT mitigates language bias in MLIR by subtracting estimated relative language vectors from document embeddings during indexing using parallel translation pairs.
CAPED reduces incidental visual privacy leakage in mobile GUI agents from 0.766 to 0.268 on seeded AndroidWorld tasks by selectively exposing only task-relevant screen content.
First end-to-end RAG on mobile NPU delivers 18.1x faster prefilling, 4x lower latency and energy than CPU on Snapdragon X Elite with equivalent quality.
Meta-study of MTEB rankings introduces dataset-composition and ranking-scheme robustness indicators and finds only a small subset of models stay consistently strong across tasks, languages, and evaluation variations.
RADAR is a geometrically grounded metric that predicts cross-domain transferability by comparing layer-wise representation trajectory distributions in foundation models.
citing papers explorer
-
SkMTEB: Slovak Massive Text Embedding Benchmark and Model Adaptation
SkMTEB is the first comprehensive text embedding benchmark for Slovak, and vocabulary-trimmed E5 adaptations achieve competitive performance with much smaller models.
-
ContinuousBench: Can Differentially Private Synthetic Text Improve Capabilities?
ContinuousBench shows non-private synthetic text transfers corpus-specific capabilities while state-of-the-art DP methods fail to do so even at ε=100.
-
Is Dimensionality a Barrier for Retrieval Models?
Dimension d = O(m^{-2} log n) nearly achieves the optimal margin m^rd(+∞, A) for retrieval embeddings, with matching lower bounds showing d = O(k log(n/k)) suffices and is necessary for m = Θ(k^{-1/2}) on k-sparse query matrices.
-
BoLT: A Benchmark to Democratize Black-box Optimization Research for Expensive LLM Tasks
BoLT is a benchmark of surrogate models fitted to real LLM experiment data that enables evaluation of Bayesian and black-box optimization methods on multi-fidelity, multi-objective, high-dimensional LLM tasks.
-
ALEE: Any-Language Evaluation of Embeddings via English-Centric Minimal Pairs
ALEE generates AMR-based English minimal pairs with fine-grained semantic shifts, translates them, and evaluates embedding models on 275+ languages to expose cross-lingual gaps linked to training data and tokenization.
-
TIGER: Inverting Transformer Gradients via Embedding-Subspace Distance Optimization
TIGER turns the low-rank attention gradient subspace into a differentiable objective for continuous embedding optimization, improving reconstruction quality and robustness over prior discrete token tests especially under noise or DP.
-
LEDGER: A Long-Context Benchmark of Corporate Annual Reports for Grounded Financial Retrieval and Extraction
LEDGER provides a corpus of 4,999 annual reports with 31 labeled KPIs and three benchmarks for page-level retrieval, needle-in-haystack lookup, and full KPI extraction from long documents.
-
Agreement in Representation Space for Open-Ended Self-Consistency
EBA clusters sampled LLM generations in representation space to estimate agreement, outperforming random selection with stable scaling and showing that central positions correlate with higher generation quality.
-
In-Context Learning for the Imputation of Public Opinion Data with Large Language Models
ICL with LLMs reduces absolute imputation error for survey data versus MICE PMM across MCAR/MAR/MNAR mechanisms and yields narrower intervals with near-nominal coverage.
-
SEA-Embedding: Open and Reproducible Text Embeddings for Southeast Asia
SEA-Embedding is a fully open text embedding pipeline for Southeast Asian languages that achieves state-of-the-art performance on the SEA-BED benchmark by analyzing data composition, training objectives, and base encoder choices.
-
TeleEmbedBench: A Multi-Corpus Embedding Benchmark for RAG in Telecommunications
TeleEmbedBench is the first multi-corpus benchmark showing LLM-based embedding models significantly outperform traditional sentence-transformers on telecommunications specifications and code for retrieval accuracy and noise robustness.
-
Teaching LLMs Human-Like Editing of Inappropriate Argumentation via Reinforcement Learning
Reinforcement learning with a multi-part reward teaches LLMs to output independent, meaning-preserving sentence edits that raise argument appropriateness close to full rewriting.
-
LongTail Driving Scenarios with Reasoning Traces: The KITScenes LongTail Dataset
KITScenes LongTail supplies multimodal driving data and multilingual expert reasoning traces to benchmark models on rare scenarios beyond basic safety metrics.
-
Spectral Tempering for Embedding Compression in Dense Passage Retrieval
Spectral Tempering derives an adaptive scaling factor γ(k) from the embedding eigenspectrum via local SNR analysis and knee-point normalization to achieve near-optimal compression without training or validation.
-
Unified Work Embeddings: Contrastive Learning of a Bidirectional Multi-task Ranker
UWE is a task-agnostic bi-encoder that uses many-to-many InfoNCE and token-level soft late interaction to achieve zero-shot ranking across unseen work-related target spaces while using far fewer parameters than Qwen3-8B and improving MAP by 4.4 points.
-
ProjAgent: Procedural Similarity Retrieval for Repository-Level Code Generation
ProjAgent introduces procedural similarity—retrieving code with matching computational logic—via LLM hidden-state projections, improving repository-level code generation to 41.14% Pass@1 on REPOCOD.
-
Graph-Native Reinforcement Learning Enables Traceable Scientific Hypothesis Generation through Conceptual Recombination
Graph-PRefLexOR fine-tunes graph-native models with GRPO to organize reasoning into phases, yielding 40-65% gains in traceable hypothesis generation and 2-3x semantic diversity on 100 materials science questions.
-
Security--Fidelity Tradeoffs: The Hidden Cost of Prompt Injection Defense
Prompt injection defenses create a security-fidelity tradeoff with no model or defense achieving both high security and high fidelity on the SecFid benchmark across 1,168 examples.
-
Jailbreaking for the Average Jane: Choosing Optimal Jailbreaks via Bandit Algorithms for Automatically Enhanced Queries
Bandit algorithms learn optimal jailbreaks from noisy exploration and, paired with complexity-enhanced queries in FrankensteinBench, achieve up to 97% attack success on 15 open-weight LLMs.
-
SHIFT: Semantic Harmonization via Index-side Feature Transformation for Multilingual Information Retrieval
SHIFT mitigates language bias in MLIR by subtracting estimated relative language vectors from document embeddings during indexing using parallel translation pairs.
-
CAPED: Context-Aware Privacy Exposure Defense for Mobile GUI Agents
CAPED reduces incidental visual privacy leakage in mobile GUI agents from 0.766 to 0.268 on seeded AndroidWorld tasks by selectively exposing only task-relevant screen content.
-
Energy-Efficient On-Device RAG on a Mobile NPU: System Design and Benchmark on Snapdragon X Elite
First end-to-end RAG on mobile NPU delivers 18.1x faster prefilling, 4x lower latency and energy than CPU on Snapdragon X Elite with equivalent quality.
-
On the Robustness of Multilingual Text Embedding Rankings Across Learning Tasks, Languages, and Benchmark Datasets
Meta-study of MTEB rankings introduces dataset-composition and ranking-scheme robustness indicators and finds only a small subset of models stay consistently strong across tasks, languages, and evaluation variations.
-
RADAR: Relative Angular Divergence Across Representations
RADAR is a geometrically grounded metric that predicts cross-domain transferability by comparing layer-wise representation trajectory distributions in foundation models.
-
Structure Retention in Embedding Spaces as a Predictor of Benchmark Performance
Embedding model performance on MTEB tasks correlates strongly with nearest-neighbor overlap and ICA magnitude differences in their embedding spaces.
-
Test-Time Compute for Frozen Embedding Models through Agentic Program Search
Agentic program search over a frozen encoder API yields retrieval programs that improve nDCG@10 on held-out tasks and unseen encoder families with no per-domain training.
-
An Annotation Scheme and Classifier for Personal Facts in Dialogue
An extended annotation scheme with new categories and attributes plus a Gemma-300M-based multi-head classifier achieves 81.6% macro F1 on personal fact classification, outperforming few-shot LLM baselines by nearly 9 points with lower compute.
-
MLAIRE: Multilingual Language-Aware Information Retrieval Evaluation Protocal
MLAIRE is a protocol that evaluates multilingual retrievers on both semantic accuracy and query-language preference using parallel passages and new metrics like LPR and Lang-nDCG, showing that standard metrics hide distinct behavioral differences among retrievers.
-
Identifier-Free Code Embedding Models for Scalable Search
A fine-tuned Qwen3-Embedding model with contrastive learning outperforms baselines on bidirectional source-to-decompiled code association and generalizes to constant-algorithm tasks.
-
Iterative Definition Refinement for Zero-Shot Classification via LLM-Based Semantic Prototype Optimization
Iterative LLM-based refinement of category definitions improves zero-shot classification performance across 13 embedding models on a new 10-category web URL benchmark.
-
Lost in Decoding? Reproducing and Stress-Testing the Look-Ahead Prior in Generative Retrieval
Reproduction confirms PAG boosts generative retrieval effectiveness, but its look-ahead planning signal collapses under intent-preserving typos and query mismatches, reverting performance to unguided decoding.
-
Differences in Text Generated by Diffusion and Autoregressive Language Models
DLMs exhibit lower n-gram entropy, higher semantic coherence, and higher semantic diversity than ARMs, primarily due to bidirectional context and remasking decoding strategies.
-
LMEB: Long-horizon Memory Embedding Benchmark
LMEB is a new benchmark that evaluates embedding models on long-horizon memory retrieval and shows this skill is largely orthogonal to traditional passage-retrieval performance.
-
A Human-Centric Framework for Data Attribution in Large Language Models
Introduces a parameter-driven framework for data attribution in LLMs that enables negotiation among creators, users, and intermediaries to meet stakeholder goals within the data economy.
-
ReasonEmbed: Enhanced Text Embeddings for Reasoning-Intensive Document Retrieval
ReasonEmbed achieves a new high of 38.1 nDCG@10 on the BRIGHT benchmark for reasoning-intensive retrieval by combining a triviality-resistant data synthesis method with dynamic per-sample training weights.
-
Quantifying and Auditing LLM Evaluation via Positive--Unlabeled Learning
A positive-unlabeled learning approach using partial optimal transport is introduced to audit and correct biases in LLM-as-a-judge systems by aligning limited human positives with unlabeled outputs in embedding space.
-
Adapting Multilingual Embedding Models to Turkish via Cross-Lingual Tokenizer Surgery and Offline Distillation
A 200M-parameter Turkish sentence embedding model is adapted from a multilingual teacher via tokenizer pruning, mean-composition initialization, and offline cosine distillation, achieving 77.55% Pearson correlation on STSbTR and 7th place on TR-MTEB.
-
Grounding Text Embeddings in Stakeholder Associations
The Stakeholder Grounding Exercise shows neural text embeddings are 19-26pp less reliable than human experts at capturing semantic distinctions, with misalignment strongly correlated to poorer clustering performance (ρ=0.9), replicated across Danish policy and US AI domains.
-
TextClusterLab: An Integrated Framework for Reliable Text Clustering Studies
TextClusterLab introduces an LLM-driven generator for synthetic text clustering datasets with tunable attributes and a suitability benchmark for evaluation.
-
To MRL or not to MRL: Text Embeddings are Robust to Truncation Without Matryoshka Learning, Except In Heavy Truncation Scenarios
Truncated embeddings from non-MRL models perform comparably to or better than MRL-trained models for most truncation levels, except heavy truncation of 80% or more.
-
ML-Embed: Inclusive and Efficient Embeddings for a Multilingual World
ML-Embed releases open multilingual embedding models trained with a new 3D-ML framework that reportedly set new MTEB records on 9 of 17 benchmarks, especially in low-resource languages.
-
AuDirector: A Self-Reflective Closed-Loop Framework for Immersive Audio Storytelling
AuDirector proposes a self-reflective closed-loop multi-agent framework with identity-aware pre-production, collaborative synthesis-correction, and human-guided refinement for coherent immersive audio storytelling.
-
NAVIS: Concurrent Search and Update with Low Position-Seeking Overhead in On-SSD Graph-Based Vector Search
NAVIS improves concurrent search and update throughput in on-SSD graph vector search by up to 2.74x for insertions and 1.37x for searches through reduced position-seeking overhead.
-
How Does Chunking Affect Retrieval-Augmented Code Completion? A Controlled Empirical Study
Function-based chunking underperforms other strategies in RAG code completion by 3.57-5.64 points, with context length as the dominant factor.
-
Linear-Time and Constant-Memory Text Embeddings Based on Recurrent Language Models
Fine-tuned recurrent models like Mamba2 produce competitive text embeddings with linear-time constant-memory inference via vertical chunking, outperforming transformers in memory use.
-
Predicting Post-Traumatic Epilepsy from Clinical Records using Large Language Model Embeddings
LLM embeddings from clinical records, fused with tabular data via gradient-boosted trees, predict post-traumatic epilepsy at AUC-ROC 0.892 and AUPRC 0.798.
-
Cross-Lingual Attention Distillation with Personality-Informed Generative Augmentation for Multilingual Personality Recognition
ADAM uses personality-guided LLM augmentation and cross-lingual attention distillation to raise balanced accuracy on multilingual personality recognition to 0.6332 on Essays and 0.7448 on Kaggle, outperforming standard BCE loss.
-
SQuTR: A Robustness Benchmark for Spoken Query to Text Retrieval under Acoustic Noise
SQuTR is a large bilingual benchmark of 37,317 synthesized spoken queries under clean/low/medium/high noise, showing that retrieval quality steadily degrades as noise increases.
-
Search-R3: Unifying Reasoning and Embedding in Large Language Models
Search-R3 trains LLMs to output search embeddings as a direct product of step-by-step reasoning via supervised pre-training and a specialized RL environment that avoids full corpus re-encoding.
-
A Comparative Study on Affective Cues in Text Embeddings Across Psychological Emotion Theories
Open-weight instruction-aware encoders capture equal or greater affective information than proprietary models at word level across emotion theories, while task-tuned and proprietary encoders perform best on sentence-level classification.